Abstract
In Chap. 7 we study best approximation problems in a general Banach space. It is well known that best approximation problems have solutions only under certain assumptions on the space X. In view of the Lau-Konjagin result these assumptions cannot be removed. On the other hand, many generic results in nonlinear functional analysis hold in any Banach space. Therefore the following natural question arises: can generic results for best approximation problems be obtained in general Banach spaces? In this chapter we answer this question in the affirmative. To this end, we consider a new framework. The main feature of this new framework is that a best approximation problem is determined by a pair consisting of a point and a closed (convex) subset of a Banach space. We consider the complete metric space of such pairs equipped with a natural complete metric and show that for most (in the sense of Baire category) pairs the corresponding best approximation problem has a unique solution. We also provide some generalizations and extensions of this result.
Access provided by Autonomous University of Puebla. Download chapter PDF
Keywords
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.
7.1 Well-Posedness and Porosity
Given a nonempty closed subset A of a Banach space (X,∥⋅∥) and a point x∈X, we consider the minimization problem
It is well known that if A is convex and X is reflexive, then problem (P) always has at least one solution. This solution is unique when X is strictly convex.
If A is merely closed but X is uniformly convex, then according to classical results of Stechkin [173] and Edelstein [59], the set of all points in X having a unique nearest point in A is G δ and dense in X. Since then there has been a lot of activity in this direction. In particular, it is known [84, 88] that the following properties are equivalent for any Banach space X:
-
(A)
X is reflexive and has a Kadec-Klee norm.
-
(B)
For each nonempty closed subset A of X, the set of points in X∖A with nearest points in A is dense in X∖A.
-
(C)
For each nonempty closed subset A of X, the set of points in X∖A with nearest points in A is generic (that is, a dense G δ subset) in X∖A.
A more recent result of De Blasi, Myjak and Papini [52] establishes well-posedness of problem (P) for a uniformly convex X, closed A and a generic x∈X.
In this connection we recall that the minimization problem (P) is said to be well posed if it has a unique solution, say a 0, and every minimizing sequence of (P) converges to a 0.
A more precise formulation of the De Blasi-Myjak-Papini result mentioned above involves the notion of porosity.
Using this terminology and denoting by F the set of all points such that the minimization problem (P) is well posed, we note that De Blasi, Myjak and Papini [52] proved, in fact, that the complement X∖F is σ-porous in X.
However, the fundamental restriction in all these results is that they hold only under certain assumptions on the space X. In view of the Lau-Konjagin result mentioned above these assumptions cannot be removed. On the other hand, many generic results in nonlinear functional analysis hold in any Banach space. Therefore the following natural question arises: can generic results for best approximation problems be obtained in general Banach spaces? In [138] we answer this question in the affirmative. In this chapter we present the results obtained in [138].
To this end, we change our point of view and consider a new framework. The main feature of this new framework is that the set A in problem (P) may also vary. In our first result (Theorem 7.3) we fix x and consider the space S(X) of all nonempty closed subsets of X equipped with an appropriate complete metric, say h. We then show that the collection of all sets A∈S(X) for which problem (P) is well posed has a σ-porous complement.
In the second result (Theorem 7.4) we consider the space of pairs S(X)×X with the metric h(A,B)+∥x−y∥, where A,B∈S(X) and x,y∈X. Once again we show that the family of all pairs (A,x)∈S(X)×X for which problem (P) is well-posed has a σ-porous complement.
In our third result (Theorem 7.5) we show that for any nonempty, separable and closed subset X 0 of X, there exists a subset \(\mathcal{F}\) of (S(X),h) with a σ-porous complement such that any \(A \in {\mathcal{F}}\) has the following property:
There exists a dense G δ subset F of X 0 such that for any x∈F, the minimization problem (P) is well posed.
In order to prove these results we now provide more information on porous sets.
Let (Y,ρ) be a metric space. We denote by B ρ (y,r) the closed ball of center y∈Y and radius r>0.
The following simple observation was made in [180].
Proposition 7.1
Let E be a subset of the metric space (Y,ρ). Assume that there exist r 0>0 and β∈(0,1) such that the following property holds:
- (P1):
-
For each x∈Y and each r∈(0,r 0], there exists z∈Y∖E such that ρ(x,z)≤r and B ρ (z,βr)∩E=∅.
Then E is porous with respect to ρ.
Proof
Let x∈Y and r∈(0,r 0]. By property (P1), there exists z∈Y∖E such that
Hence B ρ (z,βr/2)⊂B ρ (x,r)∖E and Proposition 7.1 is proved. □
As a matter of fact, property (P1) can be weakened.
Proposition 7.2
Let E be a subset of the metric space (Y,ρ). Assume that there exist r 0>0 and β∈(0,1) such that the following property holds:
- (P2):
-
For each x∈E and each r∈(0,r 0], there exists z∈Y∖E such that ρ(x,z)≤r and B ρ (z,βr)∩E=∅.
Then E is porous with respect to ρ.
Proof
We may assume that β<1/2. Let x∈Y and r∈(0,r 0]. We will show that there exists z∈Y∖E such that
If B ρ (x,r/4)∩E=∅, then (7.1) holds with z=x. Assume now that B ρ (x,r/4)∩E≠∅. Then there exists
By property (P2), there exists z∈Y∖E such that
The relations (7.2) and (7.3) imply that
Thus there indeed exists z∈Y∖E satisfying (7.1). Proposition 7.2 is now seen to follow from Proposition 7.1. □
The following definition was introduced in [180].
Assume that a set Y is equipped with two metrics ρ 1 and ρ 2 such that ρ 1(x,y)≤ρ 2(x,y) for all x,y∈Y and that the metric spaces (Y,ρ 1) and (Y,ρ 2) are complete.
We say that a set E⊂Y is porous with respect to the pair (ρ 1,ρ 2) if there exist r 0>0 and α∈(0,1) such that for each x∈E and each r∈(0,r 0], there exists z∈Y∖E such that ρ 2(z,x)≤r and \(B_{\rho_{1}}(z,\alpha r ) \cap E=\emptyset\).
Proposition 7.2 implies that if E is porous with respect to (ρ 1,ρ 2), then it is porous with respect to both ρ 1 and ρ 2.
A set E⊂Y is called σ-porous with respect to (ρ 1,ρ 2) if it is a countable union of sets which are porous with respect to (ρ 1,ρ 2).
As a matter of fact, it turns out that our results are true not only for Banach spaces, but also for all complete hyperbolic spaces.
Let (X,ρ,M) be a complete hyperbolic space. For each x∈X and each A⊂X, set
Denote by S(X) the family of all nonempty closed subsets of X. For each A,B∈S(X), define
and
It is easy to see that \(\tilde{H}\) is a metric on S(X) and that the space \((S(X),\tilde{H})\) is complete.
Fix θ∈X. For each natural number n and each A,B∈S(X), we set
and
Once again it is not difficult to see that h is a metric on S(X) and that the metric space (S(X),h) is complete. Clearly,
We equip the set S(X) with the pair of metrics \(\tilde{H}\) and h.
We now state the following three results which were obtained in [138]. Their proofs are given later in this chapter.
Theorem 7.3
Let (X,ρ,M) be a complete hyperbolic space and let \(\tilde{x} \in X\). Then there exists a set Ω⊂S(X) such that its complement S(X)∖Ω is σ-porous with respect to the pair \((h,\tilde{H})\) and such that for each A∈Ω, the following property holds:
- (C1):
-
There exists a unique \(\tilde{y} \in A\) such that \(\rho(\tilde{x},\tilde{y})= \rho(\tilde{x},A)\). Moreover, for each ε>0, there exists δ>0 such that if x∈A satisfies \(\rho(\tilde{x},x) \le\rho(\tilde{x},A)+\delta\), then \(\rho(x,\tilde{y}) \le \varepsilon \).
To state the following result we endow the Cartesian product S(X)×X with the pair of metrics d 1 and d 2 defined by
Theorem 7.4
Let (X,ρ,M) be a complete hyperbolic space. There exists a set Ω⊂S(X)×X such that its complement [S(X)×X]∖Ω is σ-porous with respect to the pair (d 1,d 2) and such that for each \((A, \tilde{x}) \in \varOmega \), the following property holds:
- (C2):
-
There exists a unique \(\tilde{y} \in A\) such that \(\rho(\tilde{x},\tilde{y})= \rho(\tilde{x},A)\). Moreover, for each ε>0, there exists δ>0 such that if z∈X satisfies \(\rho(\tilde{x},z) \le\delta\), B∈S(X) satisfies h(A,B)≤δ, and y∈B satisfies ρ(y,z)≤ρ(z,B)+δ, then \(\rho(y,\tilde{y}) \le \varepsilon \).
In classical generic results the set A was fixed and x varied in a dense G δ subset of X. In our first two results the set A is also variable. However, in our third result we show that if X 0 is a nonempty, separable and closed subset of X, then for every fixed A in a dense G δ subset of S(X) with a σ-porous complement, the set of all x∈X 0 for which problem (P) is well posed contains a dense G δ subset of X 0.
Theorem 7.5
Let (X,ρ,M) be a complete hyperbolic space. Assume that X 0 is a nonempty, separable and closed subset of X. Then there exists a set \({\mathcal{F}} \subset S(X)\) such that \(S(X) \setminus{ \mathcal{F}}\) is σ-porous with respect to the pair \((h,\tilde{H})\) and such that for each \(A \in{ \mathcal{F}}\), the following property holds:
- (C3):
-
There exists a set F⊂X 0 which is a countable intersection of open and everywhere dense subsets of X 0 with the relative topology such that for each \(\tilde{x} \in F\), there exists a unique \(\tilde{y} \in A\) for which \(\rho(\tilde{x},\tilde{y}) =\rho (\tilde{x},A)\). Moreover, if \(\{y_{i}\}_{i=1}^{\infty} \subset A\) satisfies \(\lim_{i \to\infty} \rho(\tilde{x},y_{i})=\rho(\tilde{x},A)\), then \(y_{i} \to\tilde{y}\) as i→∞.
7.2 Auxiliary Results
Let (X,ρ,M) be a complete hyperbolic space and let S(X) be the family of all nonempty closed subsets of X.
Lemma 7.6
Let A∈S(X), \(\tilde{x} \in X\) and let r,ε∈(0,1). Then there exists \(\bar{x} \in X\) such that \(\rho(\bar{x},A) \le r\) and for the set \(\tilde{A}=A\cup\{\bar{x}\}\) the following properties hold:
Proof
If \(\rho(\tilde{x},A) \le r\), then the lemma holds with \(\bar{x} =\tilde{x}\) and \(\tilde{A}=A\cup\{\tilde{x}\}\). Therefore we may restrict ourselves to the case where
Choose x 0∈A such that
There exists
such that
Set \(\tilde{A}=A\cup\{\bar{x}\}\). We have by (7.9) and (7.7),
Therefore \(\rho(\tilde{x},\bar{x})=\rho(\tilde{x}, \tilde{A})\), and if \(x \in\tilde{A}\) and \(\rho(\tilde{x},x)<\rho(\tilde{x}, \tilde{A})+r/2\), then \(x=\bar{x}\). This completes the proof of Lemma 7.6. □
Before stating our next lemma we choose, for each ε∈(0,1) and each natural number n, a number
Lemma 7.7
Let A∈S(X), \(\tilde{x} \in X\) and let r,ε∈(0,1). Suppose that n is a natural number, let
and assume that
Then there exists \(\bar{x} \in X\) such that \(\rho(\bar{x},A) \le r\) and such that the set \(\tilde{A}=A\cup\{ \bar{x}\}\) has the following two properties:
if
and
then
Proof
By Lemma 7.6, there exists \(\bar{x} \in X\) such that
and such that for the set \(\tilde{A}=A\cup\{\bar{x}\}\), equality (7.13) is true and the following property holds:
Assume that \(\tilde{y} \in X\) satisfies (7.14) and B∈S(X) satisfies (7.15). We will show that
By (7.14),
When combined with (7.13), this implies that
Relations (7.13) and (7.12) imply that
It follows from (7.5) and (7.15) that
When combined with (7.10) and (7.11), this inequality implies that
Since \(\bar{x} \in\tilde{A}\), it now follows from (7.23), (7.22) and (7.5) that \(\rho(\bar{x},B)<2^{4n+1}\alpha r\) and there exists \(\bar{y} \in X\) such that
This certainly implies (7.20), as claimed.
Assume now that z∈B satisfies (7.16). It follows from (7.16), (7.20), (7.11) and (7.10) that
Relations (7.25), (7.22) and (7.14) imply that
By (7.26), (7.14), (7.11) and (7.12),
It follows from (7.23), (7.5), (7.16) and (7.27) that
Hence there exists \(\tilde{z} \in X\) such that
By (7.14), (7.28) and (7.16) we have
It follows from this inequality, (7.20), (7.11) and (7.10) that
Thus
Using this inequality, (7.28) and (7.19), we see that \(\rho(\bar{x},\tilde{z}) \le \varepsilon /2\). Combining this fact with (7.28), (7.11) and (7.10), we conclude that
Thus (7.17) holds and Lemma 7.7 is proved. □
7.3 Proofs of Theorems 7.3–7.5
Proof of Theorem 7.3
For each integer k≥1, denote by Ω k the set of all A∈S(X) which have the following property:
- (P3):
-
There exist x A ∈X and δ A >0 such that if x∈A satisfies \(\rho(x,\tilde{x}) \le\rho(\tilde{x},A)+\delta_{A}\), then ρ(x,x A )≤1/k.
Clearly, Ω k+1⊂Ω k , k=1,2,…. Set
First we will show that S(X)∖Ω is σ-porous with respect to the pair \((h,\tilde{H})\). To meet this goal it is sufficient to show that S(X)∖Ω k is σ-porous with respect to \((h,\tilde{H})\) for all sufficiently large integers k.
There exists a natural number k 0 such that \(\rho(\theta,\tilde{x}) \le k_{0}\). Let k≥k 0 be an integer. We will show that the set S(X)∖Ω k is σ-porous with respect to \((h,\tilde{H})\). For each integer n≥k 0, set
By Lemma 7.7, the set E nk is porous with respect to \((h,\tilde{H})\) for all integers n≥k 0. Since \(S(X) \setminus \varOmega _{k}=\bigcup_{n=k_{0}}^{\infty} E_{nk}\), we conclude that S(X)∖Ω k is σ-porous with respect to \((h,\tilde{H})\). Therefore S(X)∖Ω is also σ-porous with respect to \((h,\tilde{H})\).
Let A∈Ω be given. We will show that A has property (C1). By the definition of Ω k and property (P3), for each integer k≥1, there exist x k ∈X and δ k >0 such that the following property holds:
- (P4):
-
If x∈A satisfies \(\rho(x,\tilde{x}) \le\rho(\tilde{x},A)+\delta_{k}\), then ρ(x,x k )≤1/k.
Let \(\{z_{i}\}_{i=1}^{\infty} \subset A\) be such that
Fix an integer k≥1. It follows from property (P4) that for all large enough natural numbers i,
and
Since k is an arbitrary natural number, we conclude that \(\{z_{i}\}_{i=1}^{\infty}\) is a Cauchy sequence which converges to some \(\tilde{y} \in A\). It is clear that \(\rho(\tilde{x}, \tilde{y})=\rho(\tilde{x}, A)\). If the minimizer \(\tilde{y}\) were not unique, we would be able to construct a nonconvergent minimizing sequence \(\{z_{i}\}_{i=1}^{\infty}\). Thus \(\tilde{y}\) is the unique solution to problem (P) (with \(x= \tilde{x}\)) and any sequence \(\{z_{i}\}_{i=1}^{\infty} \subset A\) satisfying (7.29) converges to \(\tilde{y}\). This completes the proof of Theorem 7.3. □
Proof of Theorem 7.4
For each integer k≥1, denote by Ω k the set of all \((A,\tilde{x}) \in S(X) \times X\) which have the following property:
- (P5):
-
There exist \(\bar{x} \in X\) and \(\bar{\delta}>0\) such that if x∈X satisfies \(\rho(x,\tilde{x}) \le\bar{\delta}\), B∈S(X) satisfies \(h(A,B) \le\bar{\delta}\), and y∈B satisfies \(\rho(y,x) \le\rho(x, B)+\bar{\delta}\), then \(\rho(y,\bar{x}) \le1/k\).
Clearly Ω k+1⊂Ω k , k=1,2,… . Set
First we will show that [S(X)×X]∖Ω is σ-porous with respect to the pair (d 1,d 2). For each pair of natural numbers n and k, set
By Lemma 7.7, the set E nk is porous with respect to (d 1,d 2) for all natural numbers n and k. Since
the set [S(X)×X]∖Ω is σ-porous with respect to (d 1,d 2), by definition.
Let \((A,\tilde{x}) \in \varOmega \). We will show that \((A,\tilde{x})\) has property (C2).
By the definition of Ω k and property (P5), for each integer k≥1, there exist x k ∈X and δ k >0 with the following property:
- (P6):
-
If x∈X satisfies \(\rho(x,\tilde{x}) \le\delta_{k}\), B∈S(X) satisfies h(A,B)≤δ k , and y∈B satisfies ρ(y,x)≤ρ(x,B)+δ k , then ρ(y,x k )≤1/k.
Let \(\{z_{i}\}_{i=1}^{\infty} \subset A\) be such that
Fix an integer k≥1. It follows from property (P6) that for all large enough natural numbers i,
and
Since k is an arbitrary natural number, we conclude that \(\{z_{i}\}_{i=1}^{\infty}\) is a Cauchy sequence which converges to some \(\tilde{y} \in A\). Clearly, \(\rho(\tilde{x}, \tilde{y})=\rho(\tilde{x}, A)\). It is not difficult to see that \(\tilde{y}\) is the unique solution to the minimization problem (P) with \(x= \tilde{x}\).
Let ε>0 be given. Choose an integer k>4/min{1,ε}. By property (P6),
Assume that z∈X satisfies \(\rho(z,\tilde{x}) \le\delta_{k}\), B∈S(X) satisfies h(A,B)≤δ k and y∈B satisfies ρ(y,z)≤ρ(z,B)+δ k . Then it follows from property (P6) that ρ(y,x k )≤1/k. When combined with (7.31), this implies that \(\rho(y,\tilde{y}) \le2/k <\varepsilon \). This completes the proof of Theorem 7.4. □
Proof of Theorem 7.5
Let \(\{x_{i}\}_{i=1}^{\infty} \subset X_{0}\) be an everywhere dense subset of X 0. For each natural number p, there exists a set \({\mathcal{F}}_{p} \subset S(X)\) such that Theorem 7.3 holds with \(\tilde{x}=x_{p}\) and \(\varOmega ={\mathcal{F}}_{p}\). Set \({\mathcal{F}}=\bigcap_{p=1}^{\infty} {\mathcal{F}}_{p}\). Clearly, \(S(X) \setminus {\mathcal{F}}\) is σ-porous with respect to the pair \((h,\tilde{H})\).
Let \(A \in{ \mathcal{F}}\) and let p≥1 be an integer. By Theorem 7.3, which holds with \(\tilde{x}=x_{p}\) and \(\varOmega ={\mathcal{F}}_{p}\), there exists a unique \(\bar{x}_{p} \in A\) such that
and the following property holds:
- (P7):
-
For each integer k≥1, there exists δ(p,k)>0 such that if x∈A satisfies ρ(x,x p )≤ρ(x p ,A)+4δ(p,k), then \(\rho(x,\bar{x}_{p}) \le1/k\).
For each pair of natural numbers p and k, set
It follows from property (P7) that for each pair of integers p,k≥1, the following property holds:
- (P8):
-
If x∈A, z∈X 0, ρ(z,x p )≤δ(p,k) and ρ(z,x)≤ρ(z,A)+δ(p,k), then \(\rho(x,\bar{x}_{p}) \le1/k\).
Set
Clearly, F is a countable intersection of open and everywhere dense subsets of X 0.
Let x∈F be given. Consider a sequence \(\{x_{i}\}_{i=1}^{\infty} \subset A\) such that
Let ε>0. Choose a natural number k>8−1/min{1,ε}. There exists an integer p≥1 such that x∈V(p,k). By the definition of V(p,k), ρ(x,x p )<δ(p,k). It follows from this inequality and property (P8) that for all sufficiently large integers i, ρ(x,x i )≤ρ(x,A)+δ(p,k) and \(\rho(x_{i},\bar{x}_{p}) \le1/k <\varepsilon /2\). Since ε is an arbitrary positive number, we conclude that \(\{x_{i}\}_{i=1}^{\infty}\) is a Cauchy sequence which converges to \(\tilde{y} \in A\). Clearly, \(\tilde{y}\) is the unique minimizer of the minimization problem z→ρ(x,z), z∈A. Note that we have shown that any sequence \(\{x_{i}\}_{i=1}^{\infty} \subset A\) satisfying (7.33) converges to \(\tilde{y}\). This completes the proof of Theorem 7.5. □
7.4 Generalized Best Approximation Problems
Given a closed subset A of a Banach space X, a point x∈X and a continuous function f:X→R 1, we consider the problem of finding a solution to the minimization problem min{f(x−y):y∈A}. For a fixed function f, we define an appropriate complete metric space \(\mathcal{M}\) of all pairs (A,x) and construct a subset Ω of \(\mathcal{M}\), which is a countable intersection of open and everywhere dense sets such that for each pair in Ω, our minimization problem is well posed.
Let (X,∥⋅∥) be a Banach space and let f:X→R 1 be a continuous function. Assume that
and that for each integer n≥1, there exists an increasing function ϕ n :(0,1)→(0,1) such that
for all x∈X satisfying ∥x∥≤n and all α∈(0,1). It is clear that (7.37) holds if f is convex.
Given a closed subset A of X and a point x∈X, we consider the minimization problem
This problem was studied by many mathematicians mostly in the case where f(x)=∥x∥. We recall that the minimization problem (P) is said to be well posed if it has a unique solution, say a 0, and every minimizing sequence of (P) converges to a 0. In other words, if \(\{y_{i}\}_{i=1}^{\infty} \subset A\) and lim i→∞ f(x−y i )=f(x−a 0), then lim i→∞ y i =a 0.
Note that in the studies of problem (P) [52, 59, 84, 88, 173], the function f is the norm of the space X. There are some additional results in the literature where either f is a Minkowski functional [51, 93] or the function ∥x−y∥, y∈A, is perturbed by some convex function [42].
However, the fundamental restriction in all these results is that they only hold under certain assumptions on either the space X or the set A. In view of the Lau-Konjagin result mentioned above, these assumptions cannot be removed. On the other hand, many generic results in nonlinear functional analysis hold in any Banach space. Therefore a natural question is whether generic existence results for best approximation problems can be obtained for general Banach spaces. Positive answers to this question in the special case where f=∥⋅∥ can be found in Sects. 7.1–7.3. In the next sections, which are based on [143], we answer this question in the affirmative for a general function f satisfying (7.34)–(7.37).
To this end, we change our point of view and consider another framework, the main feature of which is that the set A in problem (P) can also vary. We prove four theorems which were established in [143]. In our first result (Theorem 7.8), we fix x and consider the space S(X) of all nonempty closed subsets of X equipped with an appropriate complete metric, say h. We then show that the collection of all sets A∈S(X) for which problem (P) is well posed contains an everywhere dense G δ set. In the second result (Theorem 7.9), we consider the space of pairs S(X)×X with the metric h(A,B)+∥x−y∥, A,B∈S(X), x,y∈X. Once again, we show that the family of all pairs (A,x)∈S(X)×X for which problem (P) is well posed contains an everywhere dense G δ set. In our third result (Theorem 7.10), we show that for any separable closed subset X 0 of X, there exists an everywhere dense G δ subset \(\mathcal{F}\) of (S(X),h) such that any \(A \in \mathcal{F}\) has the following property: there exists a G δ dense subset F of X 0 such that for any x∈F, problem (P) is well posed.
In our fourth result (Theorem 7.11), we show that a continuous coercive convex f:X→R 1 which has a unique minimizer and a certain well-posedness property (on the whole space X) has a unique minimizer and the same well-posedness property on a generic closed subset of X.
7.5 Theorems 7.8–7.11
We recall that (X,∥⋅∥) is a Banach space, f:X→R 1 is a continuous function satisfying (7.34)–(7.36) and that for each integer n≥1, there exists an increasing function ϕ n :(0,1)→(0,1) such that (7.37) is true.
For each x∈X and each A⊂X, set
and
Denote by S(X) the collection of all nonempty closed subsets of X. For each A,B∈S(X), define
and
Here we use the convention that ∞/∞=1.
It is not difficult to see that the metric space \((S(X),\tilde{H})\) is complete.
For each natural number n and each A,B∈S(X), we set
and
Once again, it is not difficult to see that h is a metric on S(X) and that the metric space (S(X),h) is complete. Clearly, \(\tilde{H}(A,B) \ge h(A,B) \) for all A,B∈S(X).
We equip the set S(X) with the pair of metrics \(\tilde{H}\) and h. The topologies induced by the metrics \(\tilde{H}\) and h on S(X) will be called the strong topology and the weak topology, respectively.
We now state Theorems 7.8–7.11.
Theorem 7.8
Let \(\tilde{x} \in X\). Then there exists a set Ω⊂S(X), which is a countable intersection of open (in the weak topology) everywhere dense (in the strong topology) subsets of S(X), such that for each A∈Ω, the following property holds:
- (C1):
-
There exists a unique \(\tilde{y} \in A\) such that \(f(\tilde{x}- \tilde{y})= \rho_{f} (\tilde{x},A)\). Moreover, for each ε>0, there exists δ>0 such that if x∈A satisfies \(f(\tilde{x}-x) \le\rho_{f}(\tilde{x},A)+\delta\), then \(\|x-\tilde{y}\| \le \varepsilon \).
To state our second result we endow the Cartesian product S(X)×X with the pair of metrics d 1 and d 2 defined by
We will refer to the topologies induced on S(X)×X by d 2 and d 1 as the strong and weak topologies, respectively.
Theorem 7.9
There exists a set Ω⊂S(X)×X, which is a countable intersection of open (in the weak topology) everywhere dense (in the strong topology) subsets of S(X)×X, such that for each \((A, \tilde{x}) \in \varOmega \), the following property holds:
- (C2):
-
There exists a unique \(\tilde{y} \in A\) such that \(f(\tilde{x}-\tilde{y})= \rho_{f}(\tilde{x},A)\). Moreover, for each ε>0, there exists δ>0 such that if z∈X satisfies \(\|z-\tilde{x}\| \le\delta\), B∈S(X) satisfies h(A,B)≤δ, and y∈B satisfies f(z−y)≤ρ f (z,B)+δ, then \(\|y-\tilde{y}\| \le \varepsilon \).
In most classical generic results the set A was fixed and x varied in a dense G δ subset of X. In our first two results the set A is also variable. However, our third result shows that for every fixed A in a dense G δ subset of S(X), the set of all x∈X for which problem (P) is well posed contains a dense G δ subset of X.
Theorem 7.10
Assume that X 0 is a closed separable subset of X. Then there exists a set \({\mathcal{F}} \subset S(X)\), which is a countable intersection of open (in the weak topology) everywhere dense (in the strong topology) subsets of S(X), such that for each \(A \in \mathcal{F}\), the following property holds:
- (C3):
-
There exists a set F⊂X 0, which is a countable intersection of open and everywhere dense subsets of X 0 with the relative topology, such that for each \(\tilde{x} \in F\), there exists a unique \(\tilde{y} \in A\) for which \(f(\tilde{x}-\tilde{y}) =\rho_{f}(\tilde{x},A)\). Moreover, if \(\{y_{i}\}_{i=1}^{\infty} \subset A\) satisfies \(\lim_{i \to\infty} f(\tilde{x}-y_{i})=\rho_{f}(\tilde{x},A)\), then \(y_{i} \to\tilde{y}\) as i→∞.
Now we will show that Theorem 7.8 implies the following result.
Theorem 7.11
Assume that g:X→R 1 is a continuous convex function such that inf{g(x):x∈X} is attained at a unique point y ∗∈X, lim∥u∥→∞ g(u)=∞, and if \(\{y_{i}\}_{i=1}^{\infty} \subset X \) and lim i→∞ g(y i )=g(y ∗), then y i →y ∗ as i→∞. Then there exists a set Ω⊂S(X), which is a countable intersection of open (in the weak topology) everywhere dense (in the strong topology) subsets of S(X), such that for each A∈Ω, the following property holds:
- (C4):
-
There is a unique y A ∈A such that g(y A )=inf{g(y):y∈A}. Moreover, for each ε>0, there exists δ>0 such that if y∈A satisfies g(y)≤g(y A )+δ, then ∥y−y A ∥≤ε.
Proof
Define f(x)=g(−x), x∈X. Clearly, f is convex and satisfies (7.34)–(7.36). Therefore Theorem 7.8 is valid with \(\tilde{x}=0\) and there exists a set Ω⊂S(X), which is a countable intersection of open (in the weak topology) everywhere dense (in the strong topology) subsets of S(X), such that for each A∈Ω, the following property holds:
There is a unique \(\tilde{y} \in A\) such that
Moreover, for each ε>0, there exists δ>0 such that if x∈A satisfies
then \(\|x-\tilde{y}\| \le \varepsilon \). Theorem 7.11 is proved. □
It is easy to see that in the proofs of Theorems 7.8–7.10 we may assume without loss of generality that inf{f(x):x∈X}=0. It is also not difficult to see that we may assume without loss of generality that x ∗=0. Indeed, instead of the function f(⋅) we can consider f(⋅+x ∗). This new function also satisfies (7.34)–(7.37). Once Theorems 7.8–7.10 are proved for this new function, they will also hold for the original function f because the mapping (A,x)→(A,x+x ∗), (A,x)∈S(X)×A, is an isometry with respect to both metrics d 1 and d 2.
7.6 A Basic Lemma
Lemma 7.12
Let A∈S(X), \(\tilde{x} \in X\), and let r,ε∈(0,1). Then there exists \(\tilde{A} \in S(X)\), \(\bar{x} \in\tilde{A}\), and δ>0 such that
and such that the following property holds:
For each \(\tilde{y} \in X \) satisfying \(\|\tilde{y} -\tilde{x}\| \le \delta\), each B∈S(X) satisfying \(h(B,\tilde{A}) \le\delta\), and each z∈B satisfying
the inequality \(\|z-\bar{x}\| \le \varepsilon \) holds.
Proof
There are two cases: either \(\rho(\tilde{x}, A) \le r\) or \(\rho(\tilde{x}, A) > r\). Consider the first case where
Set
Clearly, (7.42) is true. Fix an integer \(n >\|\tilde{x}\|\). By (7.36), there is ξ∈(0,1) such that
Using (7.34), we choose a number δ∈(0,1) such that
and
Let
and let z∈B satisfy (7.43). By (7.49) and (7.41), \(h_{n}(\tilde{A},B)(1+h_{n}(\tilde{A},B))^{-1} \le2^{n} \delta\). This implies that \(h_{n}(\tilde{A},B)(1-2^{n}\delta) \le2^{n}\delta\). When combined with (7.47), this inequality shows that \(h_{n}(\tilde{A},B) \le2^{n+1}\delta\). Since \(n >\|\tilde{x}\| \), the last inequality, when combined with (7.44) and (7.41), implies that \(\rho(\tilde{x},B) \le2^{n+1}\delta\). Hence there is x 0∈B such that \(\|\tilde{x} -x_{0}\| \le2^{n+2}\delta\). This inequality and (7.49) imply in turn that \(\|\tilde{y} -x_{0}\| \le2^{n+3} \delta\). The definition of δ (see (7.48)) now shows that \(f(\tilde{y}-x_{0}) \le\xi\). Combining this inequality with (7.43), (7.47) and the inclusion x 0∈B, we see that
It now follows from (7.46) that \(\|z-\tilde{y}\| \le \varepsilon /2\). Hence (7.47), (7.49) and (7.45) imply that \(\|\bar{x}-z\|\le \varepsilon \). This concludes the proof of the lemma in the first case.
Now we turn our attention to the second case where
For each t∈[0,r], set
and
It is clear that μ(t), t∈[0,r], is a decreasing function. Choose a number
such that μ is continuous at t 0. By (7.35), there exists a natural number n which satisfies the following conditions:
and
Let ϕ n :(0,1)→(0,1) be an increasing function for which (7.37) is true. Choose a positive number γ∈(0,1) such that
Next, choose a positive number δ 0<1/4 such that
and
Finally, choose a vector x 0 such that
It follows from (7.62), (7.52) and (7.55) that
and hence by (6.51),
It follows from (7.62) and (7.57) that
There exist \(\bar{x} \in\{\alpha x_{0}+(1-\alpha)\tilde{x}: \alpha\in (0,1)\}\) and α 0∈(0,1) such that
and
By (7.67) and (7.66), \(r/2=\|\bar{x}-x_{0}\|=\|\alpha_{0}x_{0}+(1-\alpha_{0})\tilde{x}-x_{0}\|= (1-\alpha_{0})\|\tilde{x}-x_{0}\|\) and
Relations (7.68) and (7.65) imply that
Set
Now we will estimate \(f(\tilde{x}-\bar{x})\). By (7.67), (7.65), (7.37), (7.62) and (7.69),
Thus
By (7.70), (7.53), (7.58) and (7.71), for each \(x \in \tilde{A}\setminus\{\bar{x}\} \subset A_{t_{0}}\),
and therefore
There exists δ∈(0,δ 0) such that
and
By (7.70), (7.40), (7.66), (7.62), (7.55) and (7.52),
Relations (7.76) and (7.73) imply (7.42). Assume now that
and
First we will show that
By (7.78) and the definition of h (see (7.41)), \(h_{n}(\tilde{A},B)(1+h_{n}(\tilde{A},B))^{-1} \le2^{n}\delta\). When combined with (7.74), this inequality implies that
It follows from (7.41) and the definition of n (see (7.57), (7.56)) that \(\|\tilde{x}-\bar{x}\| \le n/2\) and \(\|\bar{x}\| \le n\). When combined with (7.70) and (7.80), this implies that \(\rho(\bar{x}, B) \le2^{n+1}\delta\). Therefore there exists \(\bar{y} \in B\) such that \(\|\bar{x}-\bar{y}\| \le2^{n+2}\delta\). Combining this inequality with (7.77), we see that \(\|(\bar{y}-\tilde{y})-(\bar{x}-\tilde{x})\|\le\|\bar{x}-\bar{y}\|+ \|\tilde{y}-\tilde{x}\|\le2^{n+3}\delta\). It follows from this inequality and (7.75) that \(f(\tilde{y} -\bar{y})\le f(\tilde{x}-\bar{x})+ \gamma/4\). By the last inequality and (7.71), \(f(\tilde{y} -\bar{y})\le\mu(t_{0})\phi_{n}(1-2r/n) +2\gamma\). This implies (7.79).
Assume now that z∈B satisfies (7.43). To complete the proof of the lemma it is sufficient to show that \(\|\bar{x}-z\| \le \varepsilon \). Assume the contrary. Then
We will show that there exists \(\bar{z} \in\tilde{A}\) such that
We have already shown that (7.80) holds. By (7.43), (7.79), (7.58) and (7.74),
Hence \(\|z-\tilde{y}\| \le n/4\) by (7.57), and by (7.77) and (7.56),
Thus ∥z∥≤n. The inclusion z∈B and (7.80) now imply that \(\rho(z,\tilde{A}) \le h_{n}(B,\tilde{A}) \le2^{n+1}\delta\). Therefore there exists \(\bar{z} \in\tilde{A}\) such that (7.82) holds. It follows from (7.82), (7.81), (7.70), (7.74) and (7.59) that
By (7.82) and (7.77), \(\|z+\tilde{x} -\tilde{y} -\bar{z}\| \le\|\tilde{x} -\tilde{y}\|+\|z-\bar{z}\| \le2^{n+2}\delta+\delta\le2^{n+3}\delta\). It follows from this inequality, (7.83), (7.52) and (7.74) that
Thus \(z+\tilde{x}-\tilde{y} \in A_{t_{0}+\delta_{0}}\). By this inclusion, (7.52), (7.53) and (7.61),
Hence, by (7.43), (7.79), (7.59) and (7.74),
Thus μ(t 0)−γ≤ϕ n (1−2r/n)μ(t 0)+3γ, which contradicts (7.58). This completes the proof of Lemma 7.12. □
7.7 Proofs of Theorems 7.8–7.11
The cornerstone of our proofs is the property established in Lemma 7.12.
By Lemma 7.12, for each (A,x)∈S(X)×X and each integer k≥1, there exist A(x,k)∈S(X), \(\bar{x}(A,k) \in A(x,k)\), and δ(x,A,k)>0 such that
and the following property holds:
- (P1):
-
For each y∈X satisfying ∥y−x∥≤2δ(x,A,k), each B∈S(X) satisfying h(B,A(x,k))≤2δ(x,A,k) and each z∈B satisfying f(y−z)≤ρ f (y,B)+2δ(x,A,k), the inequality \(\|z-\bar{x}(A,k)\| \le2^{-k}\) holds.
For each (A,x)∈S(X)×X and each integer k≥1, define
and
Now set
and for each x∈X let
It is easy to see that Ω x ×{x}⊂Ω for all x∈X, Ω x is a countable intersection of open (in the weak topology) everywhere dense (in the strong topology) subsets of S(X) for all x∈X, and Ω is a countable intersection of open (in the weak topology) everywhere dense (in the strong topology) subsets of S(X)×X.
Completion of the proof of Theorem 7.9
Let \((A,\tilde{x}) \in \varOmega \). We will show that \((A,\tilde{x})\) has property (C2). By the definition of Ω (see (7.87)), for each integer n≥1, there exist an integer k n ≥n and a pair (A n ,x n )∈S(X)×X such that
Let \(\{z_{i}\}_{i=1}^{\infty} \subset A\) be such that
Fix an integer n≥1. It follows from (7.89), (7.85) and property (P1) that for all large enough integers i,
and
Since n≥1 is arbitrary, we conclude that \(\{z_{i}\}_{i=1}^{\infty}\) is a Cauchy sequence which converges to some \(\tilde{y} \in A\). Clearly \(f(\tilde{x}-\tilde{y})=\rho_{f}(\tilde{x}, A)\). If the minimizer \(\tilde{y}\) were not unique we would be able to construct a nonconvergent minimizing sequence \(\{z_{i}\}_{i=1}^{\infty}\). Thus \(\tilde{y}\) is the unique solution to problem (P) (with \(x= \tilde{x}\)).
Let ε>0 be given. Choose an integer n>4/min{1,ε}. By property (P1), (7.89) and (7.85),
Assume that z∈X satisfies \(\|z-\tilde{x}\| \le \delta(x_{n},A_{n},k_{n})\), B∈S(X) satisfies h(A,B)≤δ(x n ,A n ,k n ), and y∈B satisfies f(z−y)≤ρ f (z,B)+δ(x n ,A n ,k n ). Then
by (7.89) and (7.85). Now it follows from property (P1) that
When combined with (7.91), this implies that
The proof of Theorem 7.9 is complete. □
Theorem 7.8 follows from Theorem 7.9 and the inclusion \(\varOmega _{\tilde{x}} \times\{\tilde{x}\} \subset \varOmega \).
Although a variant of Theorem 7.10 also follows from Theorem 7.9 by a classical result of Kuratowski and Ulam [87], the following direct proof may also be of interest.
Proof of Theorem 7.10
Let the sequence \(\{x_{i}\}_{i=1}^{\infty} \subset X_{0}\) be everywhere dense in X 0. Set \({\mathcal{F}}=\bigcap_{p=1}^{\infty} \varOmega _{x_{p}}\). Clearly, \(\mathcal{F}\) is a countable intersection of open (in the weak topology) everywhere dense (in the strong topology) subsets of S(X).
Let \(A \in{ \mathcal{F}}\) and let p,n≥1 be integers. Clearly, \(A \in \varOmega _{x_{p}}\) and by (7.88) and (7.86), there exist A n ∈S(X) and an integer k n ≥n such that
It follows from this inequality and property (P1) that the following property holds:
- (P2):
-
For each y∈X satisfying ∥y−x p ∥≤δ(x p ,A n ,k n ) and each z∈A satisfying f(y−z)≤ρ f (y,A)+2δ(x p ,A n ,k n ), the inequality \(\|z-\bar{x}_{p}(A_{n},k_{n})\| \le2^{-n}\) holds.
Set W(p,n)={z∈X 0:∥z−x p ∥<δ(x p ,A n ,k n )} and
It is clear that F is a countable intersection of open and everywhere dense subsets of X 0.
Let x∈F be given. Consider a sequence \(\{z_{i}\}_{i=1}^{\infty} \subset A\) such that
Let ε>0. Choose an integer n>8/min{1,ε}. There exists an integer p≥1 such that x∈W(p,n). By the definition of W(p,n), ∥x−x p ∥<δ(x p ,A n ,k n ). It follows from this inequality, (7.93) and property (P2) that for all sufficiently large integers i, f(x−z i )≤ρ f (x,A)+δ(x p ,A n ,k n ) and \(\|z_{i}-\bar{x}_{p}(A_{n},k_{n})\| \le2^{-n}<\varepsilon \). Since ε>0 is arbitrary, we conclude that \(\{z_{i}\}_{i=1}^{\infty}\) is a Cauchy sequence which converges to \(\tilde{y} \in A\). Clearly, \(\tilde{y}\) is the unique minimizer of the minimization problem z→f(x−z), z∈A. Note that we have shown that any sequence \(\{z_{i}\}_{i=1}^{\infty} \subset A\) satisfying (7.93) converges to \(\tilde{y}\). This completes the proof of Theorem 7.10. □
7.8 A Porosity Result in Best Approximation Theory
Let D be a nonempty compact subset of a complete hyperbolic space (X,ρ,M) and denote by S(X) the family of all nonempty closed subsets of X. We endow S(X) with a pair of natural complete metrics and show that there exists a set Ω⊂S(X) such that its complement S(X)∖Ω is σ-porous with respect to this pair of metrics and such that for each A∈Ω and each \(\tilde{x} \in D\), the following property holds: the set \(\{y \in A: \rho(\tilde{x},y)=\rho(\tilde{x},A)\}\) is nonempty and compact, and each sequence \(\{y_{i}\}_{i=1}^{\infty} \subset A\) which satisfies \(\lim_{i \to\infty} \rho(\tilde{x},y_{i})=\rho(\tilde{x},A)\) has a convergent subsequence. This result was obtained in [147].
Let (X,ρ,M) be a complete hyperbolic space. For each x∈X and each A⊂X, set
Denote by S(X) the family of all nonempty closed subsets of X. For each A,B∈S(X), define
and
Here we use the convention that ∞/∞=1. It is easy to see that \(\tilde{H}\) is a metric on S(X) and that the metric space \((S(X),\tilde{H})\) is complete.
Fix θ∈X. For each natural number n and each A,B∈S(X), we set
and
Once again, it is not difficult to see that h is a metric on S(X) and that the metric space (S(X),h) is complete. Clearly,
We equip the set S(X) with the pair of metrics \(\tilde{H}\) and h and prove the following theorem which is the main result of [147].
Theorem 7.13
Given a nonempty compact subset D of a complete hyperbolic space (X,ρ,M), there exists a set Ω⊂S(X) such that its complement S(X)∖Ω is σ-porous with respect to the pair of metrics \((h,\tilde{H})\), and such that for each A∈Ω and each \(\tilde{x} \in D\), the following property holds:
The set \(\{y \in A: \rho(\tilde{x},y)=\rho(\tilde{x},A)\}\) is nonempty and compact and each sequence \(\{y_{i}\}_{i=1}^{\infty} \subset A\) which satisfies \(\lim_{i \to\infty} \rho(\tilde{x},y_{i})=\rho(\tilde{x},A)\) has a convergent subsequence.
7.9 Two Lemmata
Let (X,ρ,M) be a complete hyperbolic space and let D be a nonempty compact subset of X. In the proof of Theorem 7.13 we will use the following two lemmata.
Lemma 7.14
Let q be a natural number, A∈S(X), ε∈(0,1), r∈(0,1], and let Q={ξ 1,…,ξ q } be a finite subset of D. Then there exists a finite set \(\{\tilde{\xi}_{1},\dots,\tilde{\xi}_{q}\} \subset X\) such that
and such that the set \(\tilde{A}:=A\cup\{\tilde{\xi}_{1},\dots,\tilde{\xi}_{q}\}\) has the following properties:
- (P3):
-
if i∈{1,…,q}, \(x \in\tilde{A}\), and \(\rho(\xi_{i},x) \le \rho(\xi_{i},\tilde{A})+\varepsilon r/4\), then
$$\rho\bigl(x,\{\tilde{\xi}_1,\dots,\tilde{\xi}_q\}\bigr) \le \varepsilon . $$
Proof
Let i∈{1,…,q}. There are two cases: (1) ρ(ξ i ,A)≤r; (2) ρ(ξ i ,A)>r. In the first case we set
In the second case, we first choose x i ∈A for which
and then choose
such that
Clearly, (7.96) holds. Consider now the set \(\tilde{A}=A \cup\{\tilde{\xi}_{1},\dots,\tilde{\xi}_{q}\}\).
Let i∈{1,…,q}. It is not difficult to see that if ρ(ξ i ,A)≤r, then the assertion of the lemma is true. Consider the case where ρ(ξ i ,A)>r. It follows from (7.99) and (7.101) that
Therefore
and if \(x \in\tilde{A}\) and \(\rho(\xi_{i},x) \le\rho(\xi_{i},\tilde{A})+r/2\), then \(x \in\{\tilde{\xi}_{1}, \dots, \tilde{\xi}_{q}\}\). This completes the proof of Lemma 7.14. □
For each ε∈(0,1) and each natural number n, choose a number
and a natural number n 0 such that
Lemma 7.15
Let n≥n 0 be a natural number, A∈S(X), ε∈(0,1), r∈(0,1], and
Assume that
Then there exist a natural number q and a finite set \(\{\tilde{\xi}_{1},\dots,\tilde{\xi}_{q}\} \subset X\) such that
and if \(\tilde{A}:=A \cup\{\tilde{\xi}_{1},\dots,\tilde{\xi}_{q}\}\), u∈D, B∈S(X),
and
then
Proof
Since D is compact, there are a natural number q and a finite subset {ξ 1,…,ξ q } of D such that
By Lemma 7.14, there exists a finite set \(\{\tilde{\xi}_{1},\dots\tilde{\xi}_{q}\} \subset X\) such that (7.106) holds, and the set \(\tilde{A}:=A\cup\{\tilde{\xi}_{1},\dots,\tilde{\xi}_{q}\}\) satisfies (7.97) and has the following property:
- (P4):
-
If i∈{1,…,q}, \(x \in\tilde{A}\), and \(\rho(\xi_{i},x) \le \rho(\xi_{i},\tilde{A})+\varepsilon r/8\), then
$$\rho\bigl(x,\{\tilde{\xi}_1,\dots,\tilde{\xi}_q\}\bigr) \le \varepsilon /2. $$
Assume that u∈D, B∈S(X), and that (7.107) holds. By (7.110), there is j∈{1,…,q} such that
We will show that
Indeed, there exists p∈{1,…,q} such that
By (7.97),
By (7.111),
When combined with (7.113), this inequality implies that
Now (7.113), (7.105) and (7.103) imply that
It follows from (7.95) and (7.107) that
and when combined with (7.104) and (7.102), this inequality yields
Since \(\tilde{\xi}_{p} \in\tilde{A}\), it follows from (7.117), (7.116) and (7.97) that \(\rho({\tilde{\xi}}_{p},B) <2^{4n+1}\alpha r\) and there exists v∈X such that
By (7.118), (7.111), (7.113) and (7.118),
Hence (7.112) is valid.
Now let (7.108) hold. Then by (7.108), (7.112) and (7.102),
Therefore (7.119) and (7.116) imply that
It follows from this inequality, (7.111) and (7.103) that
Since z∈B, it follows from (7.97) and (7.117) that
Therefore there exists \(\tilde{z} \in\tilde{A}\) such that
By (7.111), (7.120), (7.108), (7.112) and (7.102),
and
Since \(\tilde{z} \in\tilde{A}\), it follows from (7.121) and property (P4) that \(\rho(\tilde{z},\{\tilde{\xi}_{1},\dots,\tilde{\xi}_{q}\}) \le \varepsilon /2\). When combined with (7.120) and (7.102), this inequality implies that
This completes the proof of Lemma 7.15. □
7.10 Proof of Theorem 7.13
For each integer k≥1, denote by Ω k the set of all A∈S(X) which have the following property:
- (P5):
-
There exist a nonempty finite set Q⊂X and a number δ>0 such that if u∈D, x∈A and ρ(u,x)≤ρ(u,A)+δ, then ρ(x,Q)≤1/k.
It is clear that Ω k+1⊂Ω k , k=1,2,…. Set \(\varOmega =\bigcap_{k=1}^{\infty} \varOmega _{k}\).
Let k≥n 0 (see (7.103)) be an integer. We will show that S(X)∖Ω k is σ-porous with respect to the pair \((h,\tilde{H})\). For any integer n≥k, define
By Lemma 7.15, E nk is porous with respect to the pair \((h,\tilde{H})\) for all integers n≥k. Thus \(S(X)\setminus \varOmega _{k}= \bigcup_{n=k}^{\infty}E_{nk}\) is σ-porous with respect to \((h,\tilde{H})\). Hence \(S(X) \setminus \varOmega =\bigcup_{k=n_{0}}^{\infty}(S(X)\setminus \varOmega _{k})\) is also σ-porous with respect to the pair of metrics \((h,\tilde{H})\).
Let A∈Ω. Since A∈Ω k for each integer k≥1, it follows from property (P5) that for any integer k≥1, there exist a nonempty finite set Q k ⊂X and a number δ k >0 such that the following property also holds:
- (P6):
-
If u∈D, x∈A, and ρ(u,x)≤ρ(x,A)+δ k , then ρ(x,Q k )≤1/k.
Let u∈D. Consider a sequence \(\{x_{i}\}_{i=1}^{\infty} \subset A\) such that lim i→∞ ρ(u,x i )=ρ(u,D). By property (P6), for each integer k≥1, there exists a subsequence \(\{x_{i}^{(k)}\}_{i=1}^{\infty} \) of \(\{x_{i}\}_{i=1}^{\infty}\) such that the following two properties hold:
-
(i)
\(\{x_{i}^{(k+1)}\}_{i=1}^{\infty}\) is a subsequence of \(\{x_{i}^{(k)}\}_{i=1}^{\infty}\) for all integers k≥1;
-
(ii)
for any integer k≥1, \(\rho(x_{j}^{(k)},x_{s}^{(k)}) \le2/k\) for all integers j,s≥1.
These properties imply that there exists a subsequence \(\{x_{i}^{*}\}_{i=1}^{\infty} \) of \(\{x_{i}\}_{i=1}^{\infty}\) which is a Cauchy sequence. Therefore \(\{x_{i}^{*}\}_{i=1}^{\infty}\) converges to a point \(\tilde{x} \in A\) which satisfies \(\rho(\tilde{x},u)=\lim_{i \to \infty}\rho(x_{i},u)=\rho(u,D)\). This completes the proof of Theorem 7.13.
7.11 Porous Sets and Generalized Best Approximation Problems
Given a closed subset A of a Banach space X, a point x∈X and a Lipschitzian (on bounded sets) function f:X→R 1, we consider the problem of finding a solution to the minimization problem min{f(x−y):y∈A}. For a fixed function f, we define an appropriate complete metric space \(\mathcal{M}\) of all pairs (A,x) and construct a subset Ω of \(\mathcal{M}\), with a σ-porous complement \({\mathcal{M}} \setminus \varOmega \), such that for each pair in Ω, our minimization problem is well posed.
Let (X,∥⋅∥) be a Banach space and let f:X→R 1 be a Lipschitzian (on bounded sets) function. Assume that
and that for each natural number n, there exists k n >0 such that
Clearly, (7.125) holds if f is convex.
Given a closed subset A of X and a point x∈X, we consider the minimization problem
For each x∈X and each A⊂X, set
and
Denote by S(X) the collection of all nonempty closed subsets of X. For each A,B∈S(X), define
and
Here we use the convention that ∞/∞=1.
It is not difficult to see that the metric space \((S(X),\tilde{H})\) is complete.
For each natural number n and each A,B∈S(X), we set
and
Once again, it is not difficult to see that h is a metric on S(X) and that the metric space (S(X),h) is complete. Clearly, \(\tilde{H}(A,B) \ge h(A,B) \) for all A,B∈S(X).
We equip the set S(X) with the pair of metrics \(\tilde{H}\) and h. The topologies induced by the metrics \(\tilde{H}\) and h on S(X) will be called the strong topology and the weak topology, respectively.
Let A∈S(X) and \(\tilde{x} \in X\) be given. We say that the best approximation problem
is strongly well posed if there exists a unique \(\bar{x} \in A\) such that
and the following property holds:
For each ε>0, there exists δ>0 such that if z∈X satisfies \(\|z-\tilde{x}\| \le\delta\), B∈S(X) satisfies h(A,B)≤δ, and y∈B satisfies f(z−y)≤ρ f (z,B)+δ, then \(\|y-\bar{x} \| \le \varepsilon \).
We now state four results obtained in [151]. Their proofs will be given in the next sections.
Theorem 7.16
Let \(\tilde{x} \in X\) be given. Then there exists a set Ω⊂S(X) such that its complement S(X)∖Ω is σ-porous with respect to \((h,\tilde{H})\) and for each A∈Ω, the problem \(f(\tilde{x}-y) \to\min\), y∈A, is strongly well posed.
To state our second result, we endow the Cartesian product S(X)×X with the pair of metrics d 1 and d 2 defined by
We will refer to the metrics induced on S(X)×X by d 2 and d 1 as the strong and weak metrics, respectively.
Theorem 7.17
There exists a set Ω⊂S(X)×X such that its complement (S(X)×X)∖Ω is σ-porous with respect to (d 1,d 2) and for each \((A, \tilde{x}) \in \varOmega \), the minimization problem
is strongly well posed.
In most classical generic results the set A was fixed and x varied in a dense G δ subset of X. In our first two results the set A is also variable. However, our third result shows that for every fixed A in a subset of S(X) which has a σ-porous complement, the set of all x∈X for which problem (P) is strongly well posed contains a dense G δ subset of X.
Theorem 7.18
Assume that X 0 is a closed separable subset of X. Then there exists a set \({\mathcal{F}} \subset S(X)\) such that its complement \(S(X) \setminus{ \mathcal{F}}\) is σ-porous with respect to \((h,\tilde{H})\) and for each \(A \in {\mathcal{F}}\), the following property holds:
There exists a set F⊂X 0, which is a countable intersection of open and everywhere dense subsets of X 0 with the relative topology, such that for each \(\tilde{x} \in F\), the minimization problem
is strongly well posed.
Now we will show that Theorem 7.16 implies the following result.
Theorem 7.19
Assume that g:X→R 1 is a convex function which is Lipschitzian on bounded subsets of X and that inf{g(x):x∈X} is attained at a unique point y ∗∈X, lim∥u∥→∞ g(u)=∞, and if \(\{y_{i}\}_{i=1}^{\infty} \subset X \) and lim i→∞ g(y i )=g(y ∗), then y i →y ∗ as i→∞. Then there exists a set Ω⊂S(X) such that its complement S(X)∖Ω is σ-porous with respect to \((h,\tilde{H})\) and for each A∈Ω, the following property holds:
There is a unique y A ∈A such that g(y A )=inf{g(y):y∈A}. Moreover, for each ε>0, there exists δ>0 such that if y∈A satisfies g(y)≤g(y A )+δ, then ∥y−y A ∥≤ε.
Proof
Define f(x)=g(−x), x∈X. It is clear that f is convex and satisfies (7.122)–(7.126). Therefore Theorem 7.16 is valid with \(\tilde{x}=0\) and there exists a set Ω⊂S(X) such that its complement S(X)∖Ω is σ-porous with respect to \((h, \tilde{H})\) and for each A∈Ω, the following property holds:
There is a unique \(\tilde{y} \in A\) such that
Moreover, for each ε>0, there exists δ>0 such that if B∈S(X) satisfies h(A,B)≤δ and x∈B satisfies
then \(\|x-\tilde{y}\| \le \varepsilon \). Theorem 7.19 is proved. □
It is easy to see that in the proofs of Theorems 7.16–7.18 we may assume without any loss of generality that inf{f(x):x∈X}=0. It is also not difficult to see that we may assume without loss of generality that x ∗=0. Indeed, instead of the function f(⋅) we can consider f(⋅+x ∗). This new function also satisfies (7.122)–(7.126). Once Theorems 7.16–7.18 are proved for this new function, they will also hold for the original function f because the mapping (A,x)→(A,x+x ∗), (A,x)∈S(X)×X, is an isometry with respect to both metrics d 1 and d 2.
7.12 A Basic Lemma
Let m and n be two natural numbers. Choose a number
(see (7.126)). By (7.123), there exists a natural number
such that
By (7.126), there is k m >1 such that
By (7.131), there exists a positive number
such that
Finally, we choose a positive number
Lemma 7.20
Let
A∈S(X), \(\tilde{x} \in X\), r∈(0,1], and assume that
Then there exists \(\bar{x} \in X\) such that
and for the set \(\tilde{A}:=A\cup\{\bar{x}\}\), the following property holds:
If
and
then
and
Proof
First we choose \(\bar{x} \in X\). There are two cases: (1) \(\rho(\tilde{x},A) \le r/8\); (2) \(\rho(\tilde{x},A)>r/8\). If
then we set
Now consider the second case where
First, choose x 0∈A such that
and then choose
such that
Finally, set
Clearly, there is γ∈(0,1) such that
It is easy to see that in both cases (7.137) holds and
Now assume that z∈X satisfies
We will show that \(\|\bar{x}-z\| \le(2n)^{-1}\). First consider case (1). Then by (7.152), (7.144) and (7.149),
When combined with (7.133), this inequality implies that
Now consider case (2). We first estimate \(f(\tilde{x}-\bar{x})\). By (7.150) and (7.125) (with x ∗=0 and f(x ∗)=0),
By (7.136), there is z 0∈X such that
Thus (7.146), (7.132), (7.154) and (7.136) imply that
Relations (7.155) and (7.130) imply that
It follows from (7.148), (7.150) and (7.156) that
and that
Relations (7.152) and (7.158) now imply that
There are two cases:
and
Assume that (7.160) holds. Then it follows from (7.159), (7.146) and (7.160) that
Thus z∉A and by (7.152) and (7.149),
Now assume that (7.161) is true. By (7.161) and (7.152),
When combined with (7.133), (7.148) and (7.161), this estimate implies that
and
Thus in both cases,
In other words, we have shown that the following property holds:
- (P1):
-
If z∈X satisfies (7.152), then \(\|\bar{x}-z\| \le(2n)^{-1}\).
Now assume that (7.138)–(7.140) hold. By (7.136) and (7.139), we have
Relation (7.136) implies that there is z 0∈X such that
It follows from (7.128), (7.138), (7.164), (7.134) and (7.128) that
and
Inequalities (7.166), (7.134) and (7.132) imply that ρ(z 0,B)<1, and that there is \(\tilde{z}_{0} \in X\) such that
Clearly, by (7.164) and (7.167),
Let
By (7.136), (7.163), (7.164), (7.168) and (7.167),
and there is \(\bar{u} \in X\) such that
Relations (7.171), (7.170) and (7.129) imply that
Also, relations (7.172), (7.130) and (7.170) imply the following property:
- (P2):
-
If u∈L and f(l−u)≤ρ f (l,L)+2, then ∥l−u∥≤a m and ∥u∥≤∥l∥+a m ≤2a m .
Now assume that L i ∈S(X) and l i ∈X, i=1,2, satisfy
Let
By (7.174), (7.173) and property (P2),
Relations (7.174), (7.173), (7.175), (7.165) and (7.128) imply that
When combined with (7.132) and (7.134), this inequality implies that there is v∈X such that
Inequalities (7.175) and (7.176) imply that
By (7.177), (7.175), (7.173) and (7.163),
It follows from (7.176), (7.139) and (7.173) that
By (7.179), (7.178), (7.134) and the definition of k m (see (7.131)),
Inequalities (7.180) and (7.176) imply that
and
Since (7.181) holds for any u satisfying (7.174), we conclude that
This fact implies, in turn, that
By property (P2), (7.169) and (7.140),
It follows from (7.140), (7.183), (7.165) and (7.128) that
Thus there exists \(\tilde{z} \in X\) such that
By (7.136), (7.183), (7.184), (7.134) and (7.132), we have
When combined with (7.134), (7.184), (7.139), (7.140) and (7.182), this inequality implies that
Thus we see that
It follows from property (P1), (7.152), (7.185) and (7.184) that
When combined with (7.184), (7.134) and (7.132), this inequality implies that
Thus (7.142) is proved. Inequality (7.141) follows from (7.138), (7.151), (7.134) and (7.132). Thus we have shown that (7.138)–(7.140) imply (7.141) and (7.142). Lemma 7.20 is proved. □
7.13 Proofs of Theorems 7.16–7.18
We use the notations and the definitions from the previous section.
For each natural number n, denote by \({\mathcal{F}}_{n} \) the set of all (x,A)∈X×S(X) such that the following property holds:
- (P3):
-
There exist y∈A and δ>0 such that for each \(\tilde{x} \in X\) satisfying \(\|\tilde{x}-x\|\le\delta\), each B∈S(X) satisfying h(A,B)≤δ, and each z∈B satisfying \(f(\tilde{x}-z) \le\rho_{f}(\tilde{x},B)+\delta\), the inequality ∥z−y∥≤n −1 holds.
Set
Lemma 7.21
If
then the problem f(x−y)→min, y∈A, is strongly well posed.
Proof
Let \((x,A) \in{ \mathcal{F}}\) and let n be a natural number. Since \((x,A) \in{ \mathcal{F}} \subset{ \mathcal{F}}_{n}\), there exist x n ∈A and δ n >0 such that the following property holds:
- (P4):
-
For each \(\tilde{x} \in X\) satisfying \(\|\tilde{x}-x\| \le \delta_{n}\), each B∈S(X) satisfying h(A,B)≤δ n , and each z∈B satisfying \(f(\tilde{x}-z) \le\rho_{f}(\tilde{x},B)+\delta_{n}\), the inequality ∥z−x n ∥≤n −1 holds.
Suppose that
Let n be any natural number. By (7.188) and property (P4), for all sufficiently large i we have
The second inequality of (7.189) implies that \(\{z_{i}\}_{i=1}^{\infty}\) is a Cauchy sequence and there exists
Limits (7.190) and (7.188) imply that
Clearly, \(\bar{x}\) is the unique solution of the problem f(x−z)→min, z∈A. Otherwise we would be able to construct a nonconvergent sequence \(\{z_{i}\}_{i=1}^{\infty}\) satisfying (7.188). By (7.190) and (7.189),
Let ε>0 be given. Choose a natural number
Assume that
By Property (P4), ∥z−x n ∥≤1/n. When combined with (7.192) and (7.191), this inequality implies that
Thus the problem f(x−z)→min, z∈A, is strongly well posed. Lemma 7.21 is proved. □
Proof of Theorem 7.16
For each integer n≥1, set
and let
By Lemma 7.21, (7.193) and (7.194), for each A∈Ω, the problem \(f(\tilde{x}-z) \to\min\), z∈A, is strongly well posed. In order to prove the theorem, it is sufficient to show that for each natural number n, the set S(X)∖Ω n is σ-porous with respect to \((h,\tilde{H})\). To this end, let n be any natural number.
Fix a natural number
For each integer m≥m 0, define
Since
in order to prove the theorem, it is sufficient to show that for any natural number m≥m 0, the set E m ∖Ω n is porous with respect to \((h,\tilde{H})\). Let m≥m 0 be a natural number. Define
(see (7.132) and (7.134)). Let A∈S(X) and r∈(0,1]. There are two cases:
case (1), where
and case (2), where
Consider the first case.
Let
We claim that B∉E m . Assume the contrary. Then there is u∈X such that
The definition of h m (see (7.128)) and (7.200) imply that
and
When combined with (7.202), this implies that there is v∈A such that ∥u−v∥≤1/2. Together with (7.201) this inequality implies that ∥v∥≤m+1/2, a contradiction (see (7.198)). Therefore B∉E m , as claimed. Thus we have shown that
Now consider the second case. Then by Lemma 7.20, (7.195) and (7.199), there exists \(\bar{x} \in X\) such that
and such that for the set \(\tilde{A}=A \cup\{\bar{x}\}\), the following property holds:
- (P5):
-
if B∈S(X), \(h(\tilde{A},B) \le\bar{\alpha}(m+1,n)r\), \(\tilde{y} \in X\), \(\|\tilde{y}-\tilde{x}\|\le\bar{\alpha}(m+1,n)r\), and z∈B satisfies
$$f(\tilde{y}-z) \le\rho_f(\tilde{y},B)+\bar{\alpha}(m+1,n), $$then
$$\|z-\bar{x}\| \le n^{-1} \quad\hbox{and} \quad h(A,B) \le r. $$
Clearly,
Property (P5), (7.193) and the definition of \({\mathcal{F}}_{n}\) (see (P3)) imply that
Thus in both cases we have
(Note that in the first case (7.204) is true with \(\tilde{A}=A\).)
Therefore we have shown that the set E m ∖Ω n is porous with respect to \((h,\tilde{H})\). Theorem 7.16 is proved. □
Proof of Theorem 7.17
By Lemma 7.21, in order to prove the theorem, it is sufficient to show that for any natural number n, the set \((X \times S(X)) \setminus{ \mathcal{F}}_{n}\) is σ-porous in X×S(X) with respect to \((h,\tilde{H})\). To this end, let n be a natural number. For each natural number m, define
Since
in order to prove the theorem it is sufficient to show that for each natural number m, the set \(E_{m} \setminus{ \mathcal{F}}_{n}\) is porous in X×S(X) with respect to \((h, \tilde{H})\).
Let m be a natural number. Define α ∗ by (7.197). Assume that \((\tilde{x}\times A) \in X \times S(X)\) and r∈(0,1].
There are three cases:
case (1), where
case (2), where
and case (3), where
In the first case,
Next, consider the second case. In the proof of Theorem 7.16 we have shown that
and
Finally, consider the third case. Then by Lemma 7.20, there exists \(\bar{x} \in X\) such that \(\rho(\bar{x},A) \le r/8\) and such that for the set \(\tilde{A}=A\cup\{\bar{x}\}\), property (P5) holds. Clearly,
Property (P5) implies that
Hence in all three cases we have
Note that in the first and second cases, (7.210) is true with \(A=\tilde{A}\). Therefore we have shown that the set \(E_{m}\setminus{ \mathcal{F}}_{n}\) is porous with respect to (d 1,d 2). Theorem 7.17 is proved. □
Proof of Theorem 7.18
Let \(\{x_{i}\}_{i=1}^{\infty}\) be a countable dense subset of X 0. By countable dense subset of X 0. By Theorem 7.16, for each \({\mathcal{F}}_{i} \subset S(X)\) such that \(S(X) \setminus{ \mathcal{F}}_{i}\) is σ-porous in S(X) with respect to \((h,\tilde{H})\) and such that for each A∈S(X), the problem f(x i −z)→min, z∈X, is strongly well posed. Set
Clearly, \(S(X) \setminus{ \mathcal{F}}\) is a σ-porous subset of S(X) with respect to \((h,\tilde{H})\).
Let \(A \in{ \mathcal{F}}\). Assume that n and i are natural numbers. Since the problem f(x i −z)→min, z∈A, is strongly well posed, there exists a number δ in >0 and a unique \(\bar{x}_{i} \in A\) such that
and the following property holds:
- (P6):
-
if y∈X satisfies ∥y−x i ∥≤δ in , B∈S(X) satisfies h(A,B)≤δ in , and z∈B satisfies
$$ f(y-z) \le\rho_f(y,B) +\delta_{in}, $$(7.213)then \(\|z-\bar{x}_{i}\| \le(2n)^{-1}\).
Define
Clearly, F is a countable intersection of open everywhere dense subsets of X 0. Let
For each natural number q, there exist natural numbers n q ≥q and i q such that
Assume that
Let q be a natural number. Then for all sufficiently large natural numbers k,
and by property (P6) and (7.216),
This implies that \(\{y_{k}\}_{k=1}^{\infty}\) is a Cauchy sequence and there exists \(\bar{x} =\lim_{k \to\infty}y_{k}\). By (7.217), \(f(\tilde{x}-\bar{x})= \rho_{f}(\tilde{x},A)\). Clearly, \(\bar{x}\) is the unique minimizer for the problem \(f(\tilde{x}-z) \to\min\), z∈A. Otherwise, we would be able to construct a nonconvergent sequence \(\{y_{k}\}_{k=1}^{\infty}\). By (7.218),
Let ε>0 be given. Choose a natural number
Set
By (7.216), δ>0. Assume that
and
By (7.222), (7.220) and property (P6), \(\|z-\bar{x}_{i_{q}}\| \le(2q)^{-1}\). When combined with (7.219), this inequality implies that \(\|z-\bar{x}\| \le q^{-1}<\varepsilon \). This completes the proof of Theorem 7.18. □
References
Cobzas, S. (2000). Journal of Mathematical Analysis and Applications, 243, 344–356.
de Blasi, F. S., & Myjak, J. (1998). Journal of Approximation Theory, 94, 54–72.
de Blasi, F. S., Myjak, J., & Papini, P. L. (1991). Journal of the London Mathematical Society, 44, 135–142.
Edelstein, M. (1968). Journal of the London Mathematical Society, 43, 375–377.
Konjagin, S. V. (1978). Soviet Mathematics. Doklady, 19, 309–312.
Kuratowski, C., & Ulam, S. (1932). Fundamenta Mathematicae, 19, 248–251.
Lau, K. S. (1978). Indiana University Mathematics Journal, 27, 791–795.
Li, C. (2000). Journal of Approximation Theory, 107, 96–108.
Reich, S., & Zaslavski, A. J. (2001). Topological Methods in Nonlinear Analysis, 18, 395–408.
Reich, S., & Zaslavski, A. J. (2002). Nonlinear Functional Analysis and Applications, 7, 115–128.
Reich, S., & Zaslavski, A. J. (2003). Journal of Nonlinear and Convex Analysis, 4, 165–173.
Reich, S., & Zaslavski, A. J. (2004). Nonlinear Analysis Forum, 9, 135–152.
Stechkin, S. B. (1963). Revue Roumaine de Mathématiques Pures Et Appliquées, 8, 5–13.
Zaslavski, A. J. (2001). Calculus of Variations and Partial Differential Equations, 13, 265–293.
Author information
Authors and Affiliations
Rights and permissions
Copyright information
© 2014 Springer Science+Business Media New York
About this chapter
Cite this chapter
Reich, S., Zaslavski, A.J. (2014). Best Approximation. In: Genericity in Nonlinear Analysis. Developments in Mathematics, vol 34. Springer, New York, NY. https://doi.org/10.1007/978-1-4614-9533-8_7
Download citation
DOI: https://doi.org/10.1007/978-1-4614-9533-8_7
Publisher Name: Springer, New York, NY
Print ISBN: 978-1-4614-9532-1
Online ISBN: 978-1-4614-9533-8
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)